174       int32x4_t vaccSTUV = vld1q_s32(w); w = (const void*) ((const int32_t*) w + 4);  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()  local
193       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi0xOPQRSTUV), vget_high_s16(vk0xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
211       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi1xOPQRSTUV), vget_high_s16(vk1xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
229       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi2xOPQRSTUV), vget_high_s16(vk2xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
247       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi3xOPQRSTUV), vget_high_s16(vk3xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
265       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi4xOPQRSTUV), vget_high_s16(vk4xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
283       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi5xOPQRSTUV), vget_high_s16(vk5xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
301       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi6xOPQRSTUV), vget_high_s16(vk6xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
319       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi7xOPQRSTUV), vget_high_s16(vk7xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
337       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi8xOPQRSTUV), vget_high_s16(vk8xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
355       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi9xOPQRSTUV), vget_high_s16(vk9xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
373       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi10xOPQRSTUV), vget_high_s16(vk10xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
391       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi11xOPQRSTUV), vget_high_s16(vk11xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
409       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi12xOPQRSTUV), vget_high_s16(vk12xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
427       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi13xOPQRSTUV), vget_high_s16(vk13xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
445       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi14xOPQRSTUV), vget_high_s16(vk14xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
463       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi15xOPQRSTUV), vget_high_s16(vk15xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
481       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi16xOPQRSTUV), vget_high_s16(vk16xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
499       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi17xOPQRSTUV), vget_high_s16(vk17xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
517       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi18xOPQRSTUV), vget_high_s16(vk18xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
535       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi19xOPQRSTUV), vget_high_s16(vk19xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
553       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi20xOPQRSTUV), vget_high_s16(vk20xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
571       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi21xOPQRSTUV), vget_high_s16(vk21xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
589       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi22xOPQRSTUV), vget_high_s16(vk22xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
607       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi23xOPQRSTUV), vget_high_s16(vk23xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
625       vaccSTUV = vmlal_s16(vaccSTUV, vget_high_s16(vi24xOPQRSTUV), vget_high_s16(vk24xOPQRSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
634       float32x4_t vfpaccSTUV = vcvtq_f32_s32(vaccSTUV);  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
661       vaccSTUV = vreinterpretq_s32_f32(vaddq_f32(vfpaccSTUV, vmagic_bias));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
670       vaccSTUV = vqsubq_s32(vaccSTUV, vmagic_bias_less_output_zero_point);  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
676       int16x8_t vaccOPQRSTUV = vqmovn_high_s32(vqmovn_s32(vaccOPQR), vaccSTUV);  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()
685       int16x8_t vaccOPQRSTUV = vcombine_s16(vqmovn_s32(vaccOPQR), vqmovn_s32(vaccSTUV));  in xnn_qc8_dwconv_minmax_fp32_ukernel_up32x25__neon_mul16()